43
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 1, JANUARY 1985
Automated Analysis of Rodent Three-Channel
Electrocardiograms and Vectorcardiograms
HUNG T. LE, WILLIAM C. VAN ARSDEL, ARMAND M. MAKOWSKI, MEMBER,
AND
JAMES J. BAILEY
Abstract-Exposure to toxic substances often produces alterations in
heart rhythm and electrocardiographic waveform parameters such as
PR interval, ST, etc. Analysis of such physiological changes is believed
to provide sensitive indicators for revealing the degree of toxicity sustained. The purpose of this paper is to describe a system for automated
analysis of electro- and vectorcardiograms in cardiotoxicity studies on
rats. A selection of the best-suited methods of digital signal processing
is also described.
I. INTRODUCTION
I NVESTIGATORS have been using rodent electrocardiograms
(ECG) to monitor cardiotoxicity for 50 years [1]. Because
the rodent is an inexpensive model for testing, the rodent ECG
has been used increasingly by the pharmaceutical industry and
agencies such as the Food and Drug Administration for acute
and chronic toxicity trials, nutritional studies, and testing of
food additives and drugs [2].
Up until recently, the standard practice in the pharmaceutical
industry was to use a manually read, single-lead ECG and to
define premature or ectopic beats as an end point for toxicity
[3]. This practice is believed to be insufficient in determining
the level of toxicity since changes in conduction, depolarization, and repolarization patterns can reveal a level of toxicity
at a much earlier stage.
In the past, rodent ECG's have been recorded on a human
ECG chart. Due to the inappropriate frequency response (100
Hz cutoff) of the amplifiers and recording devices, true changes
as reflected by the QRS-T patterns, P-R interval, etc., may not
be determined since rodent ECG's contain frequencies which
are much higher than 100 Hz. In addition, a single-lead ECG
may not reveal the correct onsets and offsets (flducials) of
waves and/or the changes in wave patterns owing to dipole
vector forces which are orthogonal to the lead vector. No singlechannel chart can therefore provide the opportunity to construct three-space vector loops or vectorcardiograms (VCG).
These vector loops can show changes in morphology which are
often not clearly apparent in the single ECG trace; this is why
Manuscript received December 8, 1983; revised October 1, 1984.
H. T. Le was with the Laboratory of Applied Studies, Division of
Computer Research and Technology, National Institutes of Health,
Bethesda, MD 20205. He is now with the IBM Corporation, Manassas,
VA 22110.
W. C. Van Arsdel is with the Division of Drug Biology, Center for
Drugs and Biologics, Food and Drug Administration, Washington, DC
20204.
A. M. Makowski is with the Department of Electrical Engineering,
University of Maryland, College Park, MD 20740.
E. W. Pottala and J. J. Bailey are with the Medical Applications Section,
Laboratory of Applied Studies, Division of Computer Research and
Technology, National Institutes of Health, Bethesda, MD 20205.
IEEE,
ERIK W. POTTALA,
VCG's are often done as a complementary study to the routine
ECG in humans.
Another basic difficulty in the ECG/VCG analysis is created
by power line interference, random muscle noise, and baseline
wander, which can severely affect the determination of fiducials
and pattern changes. Analog methods are not well suited for
dealing with this problem, and in the case of muscle noise,
they are completely inadequate. On the other hand, digital
processing of the signal can enhance the signal-to-noise ratio
(SNR) and minimize the effect of these disturbances, thereby
improving the reliability of determining fiducials and true
pattern alterations.
Therefore, it would appear that a well-designed system for
automated ECG analysis could produce more accurate and
more sensitive indicators for cardiotoxicity, while at the same
time saving a great deal of manpower and money. Such a system would be of great benefit to investigators interested in
testing cardiotoxicity of drugs or animal models of cardiac
disease. In addition, the ability to rapidly analyze rodent ECG's
with great accuracy further encourages the use of a greater
number of animals per study, thereby improving the overall
statistical power of such studies.
The purpose of this paper is to describe a new system, which
has been successfully implemented at the Laboratory ofApplied
Studies, for the automated analysis of rodent ECG's in toxicological studies. A selection of the best suited of digital signal
processings is illustrated; this includes analog and digital filtering, baseline removal using cubic spline fitting, and signal averaging. In addition, the methods employed to determine fiducials result in reasonable and reproducible demarcation of atrial
activity, ventricular depolarization, and ventricular repolarization to the extent to which they can be separately identified in
the surface ECG.
II. DATA COLLECTION SYSTEM: DESCRIPTION
In this section, analog recordings, hardware system, analogto-digital conversion, and signal storage are described.
A. ECG Recordings
Rat ECG's are recorded on a seven-track analog tape recorder
at a speed of 3 ' in/s. The front end of the recorder consists of
an amplifier with a frequency response from 0 to 1500 Hz.
Four simultaneous channels, namely, leads I, AVF, SAVF,
and II, are recorded on tape, with the fifth channel reserved
for event code (rat number, reading time in hours, minutes,
and seconds). At the end of each rat ECG record (1-2 min in
duration), a calibration pulse of 1 mV is inscribed on all four
channels.
U.S. Government work not protected by U.S. copyright
44
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 1, JANUARY 1985
B. Hardware System
The system consists of a MAC-16 multiprogramming computer, equipped with a nine-track digital magnetic tape, an
eight-channel analog-to-digital converter, an ARDS graphic
display/keyboard, and an ASR-33 teletype.
In addition, a Honeywell system, consisting of an analog tape
recorder/reproducer, a Systron/Donner time code reader/generator, and a Systron/Donner tape search unit, is used for playing prerecorded ECG tapes during analog-to-digital conversion.
C. ADC and Signal Storage
Analog-to-digital conversion (ADC) is performed off line on
the MAC-16 which contains a 12-bit A/D converter and eight
sample and hold circuits for simultaneous data samples. Operation of the ADC system consists of initializing the multiplexed
data channel with address, block size, etc., and of issuing a
command word to the real-time clock to select the sampling
rate. To reduce high-frequency noise and aliasing (Nyquist
rate), hardware filtering is performed with a programmable
switch/filter before the A/D conversion. The low-pass filters
are all six-pole Butterworth filters.
Data values from the A/D process are stored as 16-bit integers
in three alternating buffers, with a size of 2400 words (800
words/channel). Once a buffer is filled with data, it will be
transferred onto the digital mag tape. In one A/D conversion,
22 400 data values are collected on tape for each animal.
Once the A/D process is completed, the data are than read
and stored on disk for processing on an IBM 360/370 machine.
III. SIGNAL ANALYSIS
As indicated earlier, the rat data were recorded on analog tape
at a speed of 3 3 in/s. These data are reproduced from a tape
drive at half the recording speed and are digitized at 1250 Hz.
This results in an effective sampling rate of 2500 Hz. To avoid
interference and information losses (aliasing), the analog data
are passed through a 500 Hz Butterworth low-pass filter with
an effective cutoff frequency of 1000 Hz. Note that the sampling rate is slightly higher than twice the highest frequency
component of the signal to allow for a nonperfect filter
characteristic.
A. Preprocessing
The presence of noise/interference in the recorded electrocardiograms can create large deviations from the true amplitude
and phase characteristics of the ECG signal. Among these
interference sources, it is useful to distinguish
1) baseline drift caused by rodent movement or imperfect
electrode contact,
2) muscle potential spikes which accompany body movements or muscle tremor,
3) power line interference which causes periodic high-frequency alterations of the baseline,
4) periodic low-frequency alteration of the baseline caused
by respiration.
Preprocessing is crucial to reduce interference on the ECG
signal by minimizing the above sources of noise and artifacts.
Hence, the signals are subjected to the following preprocessing
steps to allow for a precise parameter and feature extraction.
For a complete illustration, see Fig. 1.
1) Digital Filtering: A spectral analysis of the rat ECG's
indicates that it is band limited between 0 and 200 Hz. Any
frequency component in the ECG above 200 Hz is at least 30
dB down from the peak spectral component.
In the removal of high-frequency noise, the digitized data are
filtered with a low-pass windowed finite-impulse response
(FIR) having a 200 Hz cutoff frequency (see Fig. 2). This
filtering process [4] is also known as moving average since each
sampled value of the ECG is replaced by a weighted combination of the values at several adjacent sampling points, thus
reducing4the effect of random variations at any one data point.
Such filters are often implemented because of their precise linear phase characteristic; this is crucial here since the ECG signals
must be in a linear time phase relationship to each other. Indeed, implementation of a nonlinear phase filter would cause
gross distortions to the ECG waveforms. In addition, nonrecursive FIR filters are inherently stable, conceptually simple to
design, and are relatively insensitive to the effects of quantization error [4].
2) Baseline-Drift Determination, Evaluation, and Correction
a) Baseline Definition: There is some controversy as to
where to establish the baseline reference, the so-called "isoelectric baseline." However, the best choice for a reference level
seems to be the point immediately before the QRS complex
(P-R knots) since the least amount of error in baseline alterations is encountered here [5]. The reader is referred to Section
III-B1) for a method of QRS-onset detection [see Fig. l(a)].
b) Baseline Evaluation and Correction: Computer processing of electrocardiograms requires the ECG signal to have a
reasonably steady baseline since measurements of wave amplitude are made relative to the baseline. Thus, any baseline drift
resulting from rodent movements or respiration must be
redue,ed.
For low-frequency baseline wander, high-pass filtering can be
used; however, attempts to filter out higher frequency baseline
drifts would unfortunately attenuate the lower band frequency
of the signal.
The method outlined below uses a regression on a set of
fiducial points (P-R knots) in each individual channel by using
a polynomial of order three or cubic spline which estimates
the respiratory artifacts [6]. The third-.order fitting polynomial is chosen over a higher order polynomial for ease of implementation since fewer constants need to be evaluated. In
addition, the power spectral density of the cubic spline interpolators closely approximates the band-limited white noise
spectrum [7]. Simple calculations showed that implementation
of a fifth-order polynomial approximation would only improve
the frequency response by an approximate 0.5 dB, while the
number of constants to be evaluated would quadruple from
three constants in the cubic case to twelve in the quintic case.
While the above process removes higher frequency baseline
noise, it also preserves low-frequency heart information. Furthermore, such a method can also be used effectively to remove
the respiratory artifacts (heart-rate dependent) since the set of
P-R knots residing on the spline function is also directly related
to the heart rate. In effect, a "self-adjustable" high-pass filter
is being implemented.
The baseline removal process consists of estimating the baseline drift by fitting a third-order polynomial through a set of
LE
et
aL: THREE-CHANNEL ELECTROCARDIOGRAMS AND
VECTORCARDIOGRAMS
45
Start C)
Read One Data
File (X,Y, Z)
Calculate Intervals
Fir
Lowp as
Arels
aaVried Consi-
~~~~~~~~~~~~~~ably?
V
Calculate
Spatial Magnitude:
SM(I)=[X 2(I)+Y2 (I)+Z
Heart
(I)I
C2
Rhythm
Disturbed or
Intolerable
Noise Level
\/
Calculate
Spatial Velocity:
SV(I)=2*SM(I)+SM(I-1 ) .Plot
-SM(I-3)-2*SM(I-4)
ECCj
Detect QRS
and
Find peak-locations
Cl
(a)
(b)
|Beat Averaging
[Find QRS-Onsetj
ii
|
<>
Similarity Matrix
An average of 10 points
before each QRS-onset is
Si[C]
taken aa the baseline
level at the corresponding
complex. This is done
separately on each lead.
Dift
<
y
ityI
Calculate
Spatial Velocity:
V(I)-M(I)-M(1-5)
blutaied?
Find Q-Onset
&
Delete
T-beginnings
QRS-offset;
Clute
i=i+i
y
Fitting Cubic Spline
C
C
Cz on leads
X,Y,Z.
M1[X2(I)+Y2(I)+Z2(I)3
'=2
,k
J =1,2,..,k; ji
Major
Baseline
Calculate
Spatial Magnitude:
5RS-peakt
T-end.
Complex j if
ci. (0.9
P-beginning
P-end
Corrected Signal
X(I)=X(I)-C (I)
|Y(I)-Y(I)-C y (I)|- Herra,
Heart rate,
Z(I)-Y'(I)-Cz
QRS-width,
PR-segment,
C3
T & P durations.
|
Plot average beats
&
(c)
vectorloops
end
(d)
(e)
Fig. 1. Program flowchart.
predetermined fiducial points (P-R knots), and then subtracting the noise estimate from the original ECG signal. The results
of baseline noise removal are illustrated in Fig. 3.
3) Template Fornation: The extraction of parameters from
individual beats is still rather difficult, if not impossible, since
much of the ECG signal is still corrupted by noise/interference.
Furthermore, computer analysis of such a low-quality ECG
signal could result in unreliable parameter measurements. To
further improve the signal quality (signal-to-noise ratio), signal
averaging can be used, assuming that the corrupting noise is
random [8]. Since the ECG waveforms are periodic, an average
beat can be formed by summing over a number of heart cycles.
In order to obtain a "representative" beat from signal averaging,
ventricular premature beat or artifacts must be removed from
the averaging process.
To. differentiate the abnormal beats from the normal ones, a
similarity matrix Si of dimension 1 X (K - 2) is introduced,
(K - 1) being the number of heart cycles and K being the number of R peaks available, with
Si = [Cij],
j = 1, 2, .. ., K - 1;
i = 1, 2, . .. K - 1
i #j
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 1, JANUARY 1985
46
0
-39. 5367
Cs
dB
Fig. 3. Baseline correction using cubic splines; S represents the raw
data. C represents the baseline estimates. CS represents the corrected
signal (CS= S- C).
-79.0734
0
.125
Frequency normalized to sampling rate
.25
equation:
Fig. 2. Frequency response to FIR filter (Hamming window).
1
Yi K-1
K-1
Yij
j=1
and Ci0's are the cross-correlation coefficients between complexes i and j. If complex i, where i = 1, is not a "representative" or normal beat, the next complex, i.e., i = 2, will be
chosen as a reference beat. This procedure is iterated until a
majority cluster of similar complexes is obtained. Any complex
which yields a cross-correlation coefficient less than 0.90 is
then deleted from the data set. The reader is referred to
Fig. 1(d).
4) Beat Averaging: Signal averaging can be used to further
improve the signal-to-noise ratio by summing over a number of
consecutive periods. Since noise is assumed to be random [5],
[8], the noncoherent noise is added out of phase, and therefore
tends to decrease with each addition, while the coherent portions of the input ECG signal reinforce each other with each
successive addition. Fig. 4 illustrates the noise suppression in
the time and frequency domain by using signal averaging. In
this figure, the residuals were obtained by subtracting the 14th
average (i.e., average of 14 complexes) from the corresponding
(lesser) averages.
The averaging process is performed as follows.
If K is the number of R peaks available within one file of
data (5600 points/ECG lead), then there exist (K - 1) heart
cycles; here, a heart cycle is defined as the number of samples
from one peak to the next.
Before any averaging can be performed, individual heartbeats
must be properly aligned. To achieve this, critical values
within the complexes, namely, the R peaks, must be used as
synchronization points. However, due to heart rate variations,
the number of samples N1 within complex j may be different
K - 1). Therefore,
from any other complex (i = 1, 2,
elimination of a number of samples which lie within the inactive region (namely, the T-P interval) is required. The choice
of such a procedure is based on the fact that small variations in
the heart rate only affect the segment between the T wave and
thePwave [1].
Once all beats are properly aligned with an equal number N
of sample values within any one complex (i.e., N=Nj, i=
1, 2,
, K - 1), averaging can be computed by the following
-
-
,
where Yi represents the ECG amplitude measured at time t = i
1, 2, * , N), and Yij indicates the ECG amplitude measured at time t = i of complex j (i= 1,** ,N and j= 1,
K- 1).
A more efficient algorithm can be obtained as follows.
Let Mj (j = 1, * *, K) represent the peak locations, and let
K be the number of R peaks as before. Define N by the equation
(i=
-
N=min[Nj],
j=l, ,K-l
where N1 is the number of sample points in' complex j.
Then, averaging can be expressed as follows:
K-1
1
Yi IK- 1 ,YwMj+i-o)
where yi represents the average ECG amplitude at time
t=i O = 1,2, ,)N12).
-
The second half of the
average
ECG complex can be obtained
as
1
YLKK 1
K
j =2
Y(MI+I-N)
where L = N/2, * * *, N.
In this manner, elimination of a number of samples within
the inactive region is automatically achieved without having to
align the individual heartbeats.
B. Processing: Waveform Detection and
Parameter Extractions
1) QRS Detection: The QRS detection is achieved by first
finding the maximum spatial velocity of an ECG file (see Fig.
1(e) and the Appendix) and then setting a threshold at 60 percent of the maximum value; this method of waveform detection
is somewhat similar to that of Stallman and Pipberger [9]. If
the spatial velocity at any point in the data record surpasses
this threshold, a window is established at 20 and 30 ms before
LE et aL: THREE-CHANNEL ELECTROCARDIOGRAMS AND VECTORCARDIOGRAMS
I
0.01
47
(d)
0.0o
I o o 1Ia
.
-0.01
4.11
100
t
190
206
387
5S
4"
o.S
0
004
0.01
(b)
(e)
0.00! 5
-0.01
*0.11
0
I
100
1"
206
30?
4"
II
0.09
0.01
(C)c)
I
.-0.11
-.
.
.
(f)
290
397
i-
dl&A
0
._
1
100
0.5
o.o0s
_.J
-0.1t
0.25
4
0"
50
I"
49
Time Samples
505
04
0
0
-, -
*-
0.25
Freq.
0.s
~~~~~~~~~~~~~~~0
0
0.2
o.s
Freq.
Fig. 4. Noise suppression by beat averaging. Panel A shows residual voltages in raw tracing after template has been subtracted. Panel B shows residual of 7-beat average after template subtraction; panel C, the residual of 13-beat average
after template subtraction. The template is a 14-beat average. Panels D, E, and F show the power spectra of data in panels A, B, and C, respectively. Panels G, H, and I show power spectra of raw data, 7-beat average, and the 14-beat average.
Note that spectra on the latter panels retain their essential pattern relating to the signal, whereas spectra of residual noise
are reduced in middle panels by beat averaging.
and after the threshold. A search is then performed through
the spatial magnitude to determine the R-wave.peak.
2) QRS Onset and QRS Offset: The QRS onset is determined
by conducting a backward search through a 40 ms window
from the R-wave peak. The QRS onset is flagged as the point
where the spatial velocity goes below a threshold of 1.5 mV/s.
If the QRS onset is not detected within the prescribed time
window, the average beat is plotted and execution of the
algorithm is halted.
The QRS offset is precisely determined from the R-wave
peak as the point where the spatial velocity goes from negative
to positive. This fiducial point is also-defined as the beginning
of the Twave since, in rat ECG's, the S-T segment is absent [1].
3) T-Wave Detection: From the T-wave beginning (QRS offset), a search is performed over a 50 ms window for the peak
of the T wave. Once the peak is found, the T-wave end is
searched over a 70 ms window. The T-wave end is defined as
a point where the spatial velocity goes above a threshold of
- 0.25 mV/s. If no T wave is detected, a flag is set.
4) P-Wave Detection: The P wave is detected by defining a
slope threshold of 5 mV/s. Once the P wave is found, a backward search is conducted over a 20 ms window for the P-wave
beginning. Then, a forward search is performed from the threshold point for the P-wave end.
5) Parameter Extractions: With the P, QRS, and T waves
detected, various parameters such as the heart rate,PR segment,
P duration, QRS width, and T duration are calculated, and a
vector loop of the corresponding ECG tracing is constructed
on the Calcomp plotter. A sample run of the described algorithm is shown in Figs. 5 and 6.
IV. GENERAL PERFORMANCE
In the design of a system for the automated analysis of
rodent ECG's and VCG's, the degree of interindividual variability must be carefully considered. In the case of rodents, there
exists a very large range of variability in the ECG's rhythms
(200-600 beats/min). Thus, a self-adjusing procedure has
been developed to permit a general adaptation of wave-detection algorithms. Furthermore, the automated analysis of
rodent ECG's is complicated by the various amounts of interference such as power line interference, random muscle noise,
and especially baseline wander which often occurs during the
ECG recordings due to body movements.
An approach which Keiser and Meyer [5] developed is found
to be very effective in the removal of baseline wander; this
procedure uses cubic spline fitted through theP-R knots. Since
there is usually a substantial overlap of the baseline wander
spectrum and the cardiac signal spectrum, the cubic spline
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 1, JANUARY 1985
48
.5
.5
0
(a)
o
0
.25
(b) o
.5
x(n)-x(n-l)
Y
.5
.5
.5
0
(c)
.25
x(n)-2x( n-l )+x(n-2)
0
°
(d)
.5
.25
x(n-2)-8x(n-l)+8x(n+l)-x(n+2)
u
x,
.25
o
x,
.5
2x( n)+x( n-l )-x( n-3 )-2x( n-4)
S
.5
v 1
0
(e)
IL
.25
x(n)-x( n-5)
0
.5
Fig. 7. Spectral sensitivity of slope formulas.
Fig. 5. Plot of the X, Y, and Z lead after averaging. Lead II (standard
lead) is also shown. V: spatial magnitude; DIF: spatial velocity.
Markers " Y" indicate the QRS onset and offset, T end, P beginning,
and P end.
P WAVE
QRS SEGMENT
T
HEART RRTE (BEATS/MIN)
ORS-INTERVAL
(MSEC)
WRVE
435 .
18.000
PR- INTERVAL (MSEC)
20 .800
QT-INTERVAL(IMSECJ
89.600
P-OURATION( MSEC )
23 .200
T-OURATION(MISEC)
72.000
Fig. 6. Plot of vector loop and printout of timing parameters.
method has provided a reliable procedure to remove baseline
drifts while preserving the low-frequency components of the
signal. In addition, this process also adjusts for respiratory
variations in heart rate by fitting the spline function through a
set of P-R knots. This, in effect, results in a "self-adjustable"
high-pass filter. In 10 percent of the cases, however, sudden
baseline shifts due to large muscle spikes are present in the
QRS complexes, and therefore cannot be removed using the
cubic spline method. This condition is treated as an artifact
and is deleted from the data set during the averaging process.
In addition, if there are long segments of baseline wander with
a frequency exceeding one-fourth of the heart rate (0.83-2.5
Hz for 200-600 beats/min ECG record), the cubic spline
method will fail. This occurs rarely, but when it does, the
ECG is not interpretable visually or automatically.
Shown in Fig. 7(d) is the spectral sensitivity of the implemented slope formula for QRS detection. Considering that
the slope threshold is set at 60 percent of the maximum QRS
slope, this implies, for example, that high-frequency noise of
200 Hz must exceed 15 percent of the QRS amplitude to trigger a false detection (a signal-to-noise ratio of 7 :1). Fig. 8
illustrates a plot of the required noise amplitude as a fraction
of the QRS amplitude versus the noise frequency in order for a
false QRS detection to occur. In addition, such a slope formula
is also highly insensitive to low-frequency components. Thus,
only baseline "wander of exceedingly high magnitude can cause
a false QRS detection.
As shown in Fig. 4, the signal-averaging procedure has provided an efficient means of reducing the total noise power. On
the average, this procedure has improved the signal-to-noise
ratio by a factor qf four (amplitude). One limitation, however,
with the averaging method is that it requires a similarity cluster
LE et aL: THREE-CHANNEL ELECTROCARDIOGRAMS AND VECTORCARDIOGRAMS
49
REFERENCES
2. 3991
[1] R. Budden, D. K. Detweiler, and G. Zbinden, The Rat Electrocar-
Noise
Amplitude
(% of QRS
amplitude)
[2]
[3]
[4]
[5]
[6]
diogram in Pharmacology & Toxicology. London, England: Pergamon, 1981.
B. E. Osborne, "Uses and applications of electrocardiography in
toxicology: Experimental model systems in toxicology and their
significance in man," Excerpta Medica, vol. 15, pp. 85-97, 1974.
J. J. Spitzer, Myocardiallnjury. New York: Plenum, 1983.
A. V. Oppenheim and R. W. Schafer, Digital Signal Processing.
Englewood Cliffs, NJ: Prentice-Hall, 1975.
J. Wartak, Computers in Electrocardiography. Chicago, IL:
Thomas, 1970.
C. R. Meyer and H. N. Keiser, "Electrocardiogram baseline noise
estimation and removal using cubic splines and state-space computation techniques," Comput. Biomed. Res.,vol. 10, pp.459-470,
1977.
.1506
.02
.05
.Os
Normalized Noise Frequency
Fig. 8. Plot of noise amplitude as a fraction of QRS amplitude versus
normalized noise frequency (sampling rate = 2500). Shaded region:
region of correct detection.
[7] L. L. Horowitz, "The effects of spline interpolation on power
spectral density," IEEE Trans. Acoust., Speech, SignalProcessing,
vol. ASSP-22, pp. 22-27, Feb. 1974.
[8] H. T. Le, "A system for automated analysis of rodent ECG," M.S.
thesis, Dep. Elec. Eng., Univ. Maryland, College Park, 1984.
[9] F. W. Stallmann and H. V. Pipberger, "Automatic recognition of
electrocardiographic waves by digital computer," Circ. Res., vol. 9,
pp. 1138-1143, 1961.
of QRS complexes. In cases where there exist continuous
Hung T. Le was born in Saigon, Vietnam, on
May 10, 1961. HereceivedtheB.S.degreefrom
changes in QRS morphology (due to changes in conduction
the University of Virginia, Charlottesville, in
paths) or severe arrhythmia, the averaging process fails to con1982, and the M.S. degree from the University
struct a reliable QRS template. In cases of wandering atrial
of Maryland, College Park, in 1984, both in
_
electrical engineering.
pacemaker, the template may show a QRS pattern without a _na$
During his graduate studies at the University
definable P wave.
of Maryland, he held a position at the National
a iX
Thus far, this algorithm has been tested on 50 cases of rat
Institutes of Health, Bethesda, MD, conducting
research in digital signal processing with applicaECG's, and 48 cases or 96 percent were successfully analyzed.
tions to biomedical problems. He is presently
Those cases which failed consist of ECG tracings with a very with IBM, Manassas, VA,
working in acoustics development for the
low signal-to-noise ratio.
Submarine Advanced Combat System.
Mr. Le is a member of Tau Beta Pi and Eta Kappa Nu.
APPENDIX
SLOPE SPECTRAL SENSITIVITY
William C. Van Arsdel was born in Indianapolis,
IN, on April 27, 1920. He received the B.S. deThe reliability of fiducial determinations depends on accurate
gree
in pharmacology from the University of
QRS detections. Thus, it is crucial that an effective algorithm
Medical School, Eugene, and the Ph.D.
Oregon
m
l
>
be used to distinguish between high-frequency QRS and noise
degree in physiology (zoology) from Oregon
a | Eg State University, Corvallis.
components. Fig. 7 (a)-(e) illustrates the spectral sensitivity
He has been a Pharmacologist in the Bureau of
of different slope formulas.
Drugs (now the Center for Drugs and Biologics),
While the first difference [Fig. 7(a)] measurement is an
FDA, Washington, DC, since 1963. His thesis
and postdoctoral research at Oregon State ineffective algorithm for noise-free signal, it is very vulnerable to
volved
of large animals (cathigh-frequency noise. In the ECG algorithm, we use slope tle). His current research withelectrocardiology
the Division of Drug Biology involves
formula 7(d) in detecting the QRS complexes since it is less the use of laboratory animal indicators of cardiac toxicity and/or injury
susceptible to high-frequency noise as compared to the others. in drug testing.
However, once the average beat with a low level of high-frea
quency noise is formed, the slope formula(s) is used to detect
ag>jgiirwgi
Z Armand M. Makowski (M'82) was born in
the P, QRS, and T waves.
Watermael-Boitsfort, Belgium, on October 9,
1953. He received the Licence en Sciences
Shown in Fig. 8 is a plot of the noise as a fraction of the
a
Mathematiques from the University Libre de
QRS amplitude versus normalized frequency. In this illustra_
Bruxelles in 1975, the M.S. degree in engineera g
tion, it is assumed that the QRS frequency is centered at 100 a@|i___ gg ing-systems science from the University of CaliLos Angeles, in 1976, and the Ph.D.
Hz, and thus has a spectral sensitivity of SQ given in Fig. 7(d). adg ga afornia,
degree in applied mathematics from the UniverConsidering that the slope threshold is set at 60 percent of the
sity of Kentucky, Lexington, in 1981.
In August 1981 he joined the Faculty of the
maximum QRS amplitude, an approximate region of correct
Department
of Electrical Engineering, University
QRS detection can be evaluated as follows:
of Maryland, College Park, where he is presently an Assistant Professor.
EN)<0.6S(fN)SQ
where fN = flf, is the normalized frequency and fs = 2500 Hz
is the sampling rate. The above simplified model and assumptions yield a rough estimate of the overall performance of the
QRS detector with the presence of high-frequency noise.
His research interests lie primarily in the areas of stochastic dynamical
optimization, statistical signal processing, and more broadly, in the
applications of the theory of stochastic processes to engineering problems. More recently, he has_been working on problems of adaptive
optimal control for queueing systems, and on various applications of
nonlinear (stochastic) filtering and digital signal processing.
Dr. Makowski was a C.R.B. Graduate Fellow of the Belgian-American
Educational Foundation for the academic year 1975-1976. He is a
recipient of the 1984 NSF Presidential Young Investigator Award.
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 1, JANUARY 1985
Erik W. Pottala was born in Fitchburg, MA, in
James J. Bailey was born in Great Falls, MT, in
1939. He received the B.S. degree from Wor1934. He received the A.B. degree from Harvard
cester Polytechnic Institute, Worcester, MA, in
College, Cambridge, MA, in 1956, and the M.S.
1961, the Master's of Engineering degree from
anid M.D. degrees from the University of ColoYale University, New Haven, CT, in 1963, and
rado, Boulder, in 1961 and 1965, respectively.
the Ph.D. degree from the University of Mary_ He is the Chief of the Medical Applications
land, College Park, in 1970.
ion, Laboratory of Applied Studies in the
He is an Electronic Engineer with the Medical
Division of Computeroesearch and Technology,
National Institutes of Healtha Bethesda, MD.
Applications Section, Laboratory of Applied
E
,,
|I
Studies in the Division of Computer Research
He has worked on computer analysis of electroand Technology, National Institutes of Health,
cardiograms since 1969.
Bethesda, MD. He is currently engaged in analog-to-digital conversion
DL Bailey is currently a Co-Chairman of the American Heart Associaand physiological modeling.
tion Subcommittee on Computers and Electrocardiography.
Automated Spectral Characterization of Wheezing
in Asthmatic Children
T. RICHARD FENTON,
MEMBER, IEEE,
HANS PASTERKAMP, A. TAL,
Abstract-Breath sounds were recorded in normal and asthmatic
children over the chest and trachea. The power spectra of the sounds
were analyzed for peaks of high amplitude and high frequency as indications of wheezing. The percent of inspiration and expiration spent
wheezing was used as an indication of the severity of bronchial obstruction. Wheezing was found to be strongly dependent upon air flow, and
generally followed the changes in pulmonary function as indicated by
the forced expiratory volume at 1 s (FEV1). The trachea was found to
be a better location for analyzing wheezes than the lung.
I. INTRODUCTION
W
->,,rHEEZES
are clinically defined as abnormal, more or less
v melodic tones of distinguishable pitch which last longer
than 100 ms [141. They have been termed continuous added
lung sounds, implying that they are superimposed on normal
breath sounds and relatively continuous when compared to
other abnormal sounds, such as crackles which typically last
10 ms [16]. When studied with spectral analysis, wheezes are
seen as narrow peaks in the power spectrum, generally below
2000 Hz [8].
Wheezing is of very practical importance in the diagnosis and
management of a number of pulmonary pathologies, such as
asthma and bronchiolitis. The presence and location of the
wheeze, its duration, and its relation to the respiratory cycle
all assist the physician in his assessment of the patient [8] .
Manuscript received January 24, 1984; revised July 15, 1984. This
work was supported by the Winnipeg Children's Hospital Research
Foundation, Winnipeg, Man., Canada. The work of H. Pasterkamp
was supported by the Manitoba Lung Association, Canada. The work
of A. Tal was supported by the Faculty Fund, Faculty of Medicine,
University of Manitoba, Winnipeg, Man., Canada.
T. R. Fenton is with the Department of Biomedical Engineering,
Health Sciences Centre and the Departments of Energy and Electrical
Engineering, University of Manitoba, Winnipeg, Man., Canada R3E OZ3.
H. Pasterkamp and V. Chernick are with the Department ofPediatrics,
Uniiversity of Manitoba, Winnipeg, Man., Canada R3E OZ3.
A. Tal is with the Department of Pediatrics, University of Manitoba,
Winnipeg, Man., Canada, on leave from Ben Gurion University, BeerSheva, Israel.
AND
VICTOR CHERNICK
Simple auscultation is not sufficiently objective to permit
reliable comparison of wheezing over periods of time nor over
different parts of the chest and trachea [201. So far, few automated methods have been developed to make the detection
and characterization of wheezing more reliable and specific,
and clinical application of these techniques has been very
modest. This is due in part to the complexity of the apparatus
and to the awkwardness of the procedure relative to conventional auscultation which has remained essentially unmodified
since it was established in 1819 [12].
The most straightforward automated method is to compute
the power spectra of random finite record lengths of lung
sounds [2], [17]. These spectra are commonly described by
their predominant frequency and signal bandwidth [4], [9].
Since air flow is known to determine breath sound intensity
[131, it is normally necessary to incorporate additional calculations such as successive averaging into these analyses to achieve
reliability. If the patients are cooperative, air flow may be
included with the measurements so that the records are taken
under more defined and repeatable conditions [4]. In very
young children, respiration signals may be derived from thoracic
impedance [18].
The present study was designed to develop an objective
computer-aided analysis of wheezing in children, and to apply
the technique in studying the evolution of asthmatic attacks.
Both lung and tracheal sounds were included, as it has been
suggested that in some patients tracheal auscultation may be
more useful [14].
II. METHODS
Protocol
Five asthmatic subjects, ages 10-16 years, were studied during
spontaneous or induced asthma attacks, and 20 min after
therapeutic relief of bronchial obstruction with 200 jig of
inhaled Salbutamol. For induced asthma, three subjects were
exercised on a bicycle ergometer for approximately 10 min to
0018-9294/85/0100-0050$01.00 © 1985 IEEE